AI & Systems · Autonomous Driving

The map that thinks ahead.

Sensors see geometry. The Map Knowledge Graph adds judgment.

An interactive autonomous-driving demonstrator where the HD map becomes a queryable Knowledge Graph — lanes, signals, rules, occlusions, and priors linked into decisions the vehicle can defend, edge by edge. Pick a scenario, scrub each reasoning step, and inspect the exact evidence behind every call.

Launch the simulatorFour reasoning patterns
7Scenarios
4Reasoning patterns
25Node types
26Relation types
17HD map layers
Animated scene: at a green right turn, a parked van hides a crosswalk entry. The Knowledge Graph connects the signal, crosswalk, occlusion, and pedestrian prior into a creep-and-verify decision.EGO VEHICLEPLANNED PATHSIGNAL · GREENCROSSWALKPARKED VANFAR ENTRY · UNVERIFIEDKNOWLEDGE LAYERSENSOR GEOMETRYPEDESTRIAN PRIOR ↑SCHOOL ZONE · 15:00YIELD RULECROSSWALK PROTECTEDDECISIONCREEP · VERIFY CROSSWALK
Loading simulator...

Reasoning patterns

Four ways knowledge outranks geometry

Each scenario isolates one reasoning pattern an autonomous vehicle needs when sensors alone cannot answer the question. Open one in Knowledge Graph View and scrub the steps.

Occlusion & priors

Blind-Corner Crosswalk Risk

The signal is green and no pedestrian is visible. The Knowledge Graph knows the turn path crosses a crosswalk whose far entry hides behind a parked van — inside a pedestrian-priority zone, during an elevated-likelihood window.

Creep at 5 km/h · verify · prepare to stop

Open scenario →
Topology over signal

Green Light, No Exit

A green light says go; lane topology says there is nowhere to go. Storage math — available exit space versus vehicle length — outranks the signal, keeping the box junction clear.

Hold before the intersection despite green

Open scenario →
Legality vs executability

Left-Turn Conflict Window

The map says the left turn is legal. The scene says opposing traffic closes the conflict zone faster than the maneuver clears it. Legality comes from the map; executability comes from time-to-conflict.

Wait · yield to opposing traffic and crosswalk

Open scenario →
ODD eligibility

ODD Readiness Gate

The driver requests autonomy. Before activating, the Knowledge Graph audits the road itself: divider, lane connectivity, speed-limit confidence, signal coverage, road class, map freshness. One failed requirement gates the feature.

Do not activate · downgrade automation

Open scenario →

The knowledge substrate

Inside the Map Knowledge Graph

The HD map here is not geometry to localize against — it is a queryable body of knowledge sitting between perception and planning. Five layers turn a static map into a reasoning substrate.

  1. 01

    HD map substrate

    Lanes, connectivity, signals, crosswalks, and regulated zones as first-class objects — seventeen rendered layers, from road surface to risk overlays.

  2. 02

    Semantic ontology

    Twenty-five node types and twenty-six relation types cover the vocabulary of urban driving: maneuvers, conflict zones, occlusion risks, priors, ODD requirements.

  3. 03

    Scene fusion

    Live actors and signal states bind onto the static graph. A detected van stops being a box in a point cloud and becomes an occluder of a mapped crosswalk entry.

  4. 04

    Rule & risk evaluation

    Keep-clear storage checks, time-to-conflict windows, pedestrian-likelihood factors, and ODD audits evaluate over the joined graph — deterministically, scene after scene.

  5. 05

    Evidence-cited decisions

    Every verdict carries the exact chain of graph relations that produced it — traceable hop by hop, replayable for any scene.

Map as a sensor

Perception is bounded by line of sight and the current frame. The Knowledge Graph contributes what the city already knows: the entry behind the van, the rule in force, the risk that recurs at 15:00.

Deterministic & auditable

The same scene always yields the same verdict, and every constraint cites its evidence — the property incident review and safety cases actually need.

ODD as a query

Feature eligibility — divider presence, lane connectivity, signal coverage, map freshness — is computed from the graph on demand, not hard-coded per road.

Outlook

Where a living map graph goes next

The demonstrator isolates a pattern driving automation is converging on: explicit, queryable world knowledge between perception and planning. Three directions the same substrate scales.

Fleet-learned priors

Every vehicle that clears an occluded corner is a measurement. Aggregated across a fleet, those observations sharpen pedestrian priors, parking patterns, and occlusion maps — knowledge no single sensor set can capture.

Living map governance

Freshness, confidence, and change detection live as properties of the graph. When the world shifts — new construction, a repainted lane — eligibility and behavior degrade per edge, not per software release.

Explainability for the safety case

Regulators and review boards ask why. An evidence-cited decision trace turns 'the vehicle chose to brake' into a replayable chain of map facts, rules, and measurements.

Back to the simulatorMore projects